1 /*- 2 * Copyright (c) 1982, 1986, 1990, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95 39 * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $ 40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.91 2008/09/09 04:06:13 dillon Exp $ 41 */ 42 43 #include "opt_ktrace.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/proc.h> 48 #include <sys/kernel.h> 49 #include <sys/signalvar.h> 50 #include <sys/resourcevar.h> 51 #include <sys/vmmeter.h> 52 #include <sys/sysctl.h> 53 #include <sys/lock.h> 54 #include <sys/uio.h> 55 #ifdef KTRACE 56 #include <sys/ktrace.h> 57 #endif 58 #include <sys/xwait.h> 59 #include <sys/ktr.h> 60 #include <sys/serialize.h> 61 62 #include <sys/signal2.h> 63 #include <sys/thread2.h> 64 #include <sys/spinlock2.h> 65 #include <sys/mutex2.h> 66 67 #include <machine/cpu.h> 68 #include <machine/smp.h> 69 70 TAILQ_HEAD(tslpque, thread); 71 72 static void sched_setup (void *dummy); 73 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 74 75 int hogticks; 76 int lbolt; 77 int lbolt_syncer; 78 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 79 int ncpus; 80 int ncpus2, ncpus2_shift, ncpus2_mask; /* note: mask not cpumask_t */ 81 int ncpus_fit, ncpus_fit_mask; /* note: mask not cpumask_t */ 82 int safepri; 83 int tsleep_now_works; 84 int tsleep_crypto_dump = 0; 85 86 static struct callout loadav_callout; 87 static struct callout schedcpu_callout; 88 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 89 90 #define __DEALL(ident) __DEQUALIFY(void *, ident) 91 92 #if !defined(KTR_TSLEEP) 93 #define KTR_TSLEEP KTR_ALL 94 #endif 95 KTR_INFO_MASTER(tsleep); 96 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_beg, 0, "tsleep enter %p", sizeof(void *)); 97 KTR_INFO(KTR_TSLEEP, tsleep, tsleep_end, 1, "tsleep exit", 0); 98 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_beg, 2, "wakeup enter %p", sizeof(void *)); 99 KTR_INFO(KTR_TSLEEP, tsleep, wakeup_end, 3, "wakeup exit", 0); 100 KTR_INFO(KTR_TSLEEP, tsleep, ilockfail, 4, "interlock failed %p", sizeof(void *)); 101 102 #define logtsleep1(name) KTR_LOG(tsleep_ ## name) 103 #define logtsleep2(name, val) KTR_LOG(tsleep_ ## name, val) 104 105 struct loadavg averunnable = 106 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 107 /* 108 * Constants for averages over 1, 5, and 15 minutes 109 * when sampling at 5 second intervals. 110 */ 111 static fixpt_t cexp[3] = { 112 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 113 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 114 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 115 }; 116 117 static void endtsleep (void *); 118 static void loadav (void *arg); 119 static void schedcpu (void *arg); 120 #ifdef SMP 121 static void tsleep_wakeup_remote(struct thread *td); 122 #endif 123 124 /* 125 * Adjust the scheduler quantum. The quantum is specified in microseconds. 126 * Note that 'tick' is in microseconds per tick. 127 */ 128 static int 129 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 130 { 131 int error, new_val; 132 133 new_val = sched_quantum * ustick; 134 error = sysctl_handle_int(oidp, &new_val, 0, req); 135 if (error != 0 || req->newptr == NULL) 136 return (error); 137 if (new_val < ustick) 138 return (EINVAL); 139 sched_quantum = new_val / ustick; 140 hogticks = 2 * sched_quantum; 141 return (0); 142 } 143 144 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 145 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 146 147 /* 148 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 149 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 150 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 151 * 152 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 153 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 154 * 155 * If you don't want to bother with the faster/more-accurate formula, you 156 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 157 * (more general) method of calculating the %age of CPU used by a process. 158 * 159 * decay 95% of `lwp_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 160 */ 161 #define CCPU_SHIFT 11 162 163 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 164 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 165 166 /* 167 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 168 */ 169 int fscale __unused = FSCALE; /* exported to systat */ 170 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 171 172 /* 173 * Recompute process priorities, once a second. 174 * 175 * Since the userland schedulers are typically event oriented, if the 176 * estcpu calculation at wakeup() time is not sufficient to make a 177 * process runnable relative to other processes in the system we have 178 * a 1-second recalc to help out. 179 * 180 * This code also allows us to store sysclock_t data in the process structure 181 * without fear of an overrun, since sysclock_t are guarenteed to hold 182 * several seconds worth of count. 183 * 184 * WARNING! callouts can preempt normal threads. However, they will not 185 * preempt a thread holding a spinlock so we *can* safely use spinlocks. 186 */ 187 static int schedcpu_stats(struct proc *p, void *data __unused); 188 static int schedcpu_resource(struct proc *p, void *data __unused); 189 190 static void 191 schedcpu(void *arg) 192 { 193 allproc_scan(schedcpu_stats, NULL); 194 allproc_scan(schedcpu_resource, NULL); 195 wakeup((caddr_t)&lbolt); 196 wakeup((caddr_t)&lbolt_syncer); 197 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 198 } 199 200 /* 201 * General process statistics once a second 202 */ 203 static int 204 schedcpu_stats(struct proc *p, void *data __unused) 205 { 206 struct lwp *lp; 207 208 /* 209 * Threads may not be completely set up if process in SIDL state. 210 */ 211 if (p->p_stat == SIDL) 212 return(0); 213 214 PHOLD(p); 215 if (lwkt_trytoken(&p->p_token) == FALSE) { 216 PRELE(p); 217 return(0); 218 } 219 220 p->p_swtime++; 221 FOREACH_LWP_IN_PROC(lp, p) { 222 if (lp->lwp_stat == LSSLEEP) 223 lp->lwp_slptime++; 224 225 /* 226 * Only recalculate processes that are active or have slept 227 * less then 2 seconds. The schedulers understand this. 228 */ 229 if (lp->lwp_slptime <= 1) { 230 p->p_usched->recalculate(lp); 231 } else { 232 lp->lwp_pctcpu = (lp->lwp_pctcpu * ccpu) >> FSHIFT; 233 } 234 } 235 lwkt_reltoken(&p->p_token); 236 lwkt_yield(); 237 PRELE(p); 238 return(0); 239 } 240 241 /* 242 * Resource checks. XXX break out since ksignal/killproc can block, 243 * limiting us to one process killed per second. There is probably 244 * a better way. 245 */ 246 static int 247 schedcpu_resource(struct proc *p, void *data __unused) 248 { 249 u_int64_t ttime; 250 struct lwp *lp; 251 252 if (p->p_stat == SIDL) 253 return(0); 254 255 PHOLD(p); 256 if (lwkt_trytoken(&p->p_token) == FALSE) { 257 PRELE(p); 258 return(0); 259 } 260 261 if (p->p_stat == SZOMB || p->p_limit == NULL) { 262 lwkt_reltoken(&p->p_token); 263 PRELE(p); 264 return(0); 265 } 266 267 ttime = 0; 268 FOREACH_LWP_IN_PROC(lp, p) { 269 /* 270 * We may have caught an lp in the middle of being 271 * created, lwp_thread can be NULL. 272 */ 273 if (lp->lwp_thread) { 274 ttime += lp->lwp_thread->td_sticks; 275 ttime += lp->lwp_thread->td_uticks; 276 } 277 } 278 279 switch(plimit_testcpulimit(p->p_limit, ttime)) { 280 case PLIMIT_TESTCPU_KILL: 281 killproc(p, "exceeded maximum CPU limit"); 282 break; 283 case PLIMIT_TESTCPU_XCPU: 284 if ((p->p_flag & P_XCPU) == 0) { 285 p->p_flag |= P_XCPU; 286 ksignal(p, SIGXCPU); 287 } 288 break; 289 default: 290 break; 291 } 292 lwkt_reltoken(&p->p_token); 293 lwkt_yield(); 294 PRELE(p); 295 return(0); 296 } 297 298 /* 299 * This is only used by ps. Generate a cpu percentage use over 300 * a period of one second. 301 * 302 * MPSAFE 303 */ 304 void 305 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 306 { 307 fixpt_t acc; 308 int remticks; 309 310 acc = (cpticks << FSHIFT) / ttlticks; 311 if (ttlticks >= ESTCPUFREQ) { 312 lp->lwp_pctcpu = acc; 313 } else { 314 remticks = ESTCPUFREQ - ttlticks; 315 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 316 ESTCPUFREQ; 317 } 318 } 319 320 /* 321 * tsleep/wakeup hash table parameters. Try to find the sweet spot for 322 * like addresses being slept on. 323 */ 324 #define TABLESIZE 4001 325 #define LOOKUP(x) (((u_int)(uintptr_t)(x)) % TABLESIZE) 326 327 static cpumask_t slpque_cpumasks[TABLESIZE]; 328 329 /* 330 * General scheduler initialization. We force a reschedule 25 times 331 * a second by default. Note that cpu0 is initialized in early boot and 332 * cannot make any high level calls. 333 * 334 * Each cpu has its own sleep queue. 335 */ 336 void 337 sleep_gdinit(globaldata_t gd) 338 { 339 static struct tslpque slpque_cpu0[TABLESIZE]; 340 int i; 341 342 if (gd->gd_cpuid == 0) { 343 sched_quantum = (hz + 24) / 25; 344 hogticks = 2 * sched_quantum; 345 346 gd->gd_tsleep_hash = slpque_cpu0; 347 } else { 348 gd->gd_tsleep_hash = kmalloc(sizeof(slpque_cpu0), 349 M_TSLEEP, M_WAITOK | M_ZERO); 350 } 351 for (i = 0; i < TABLESIZE; ++i) 352 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 353 } 354 355 /* 356 * This is a dandy function that allows us to interlock tsleep/wakeup 357 * operations with unspecified upper level locks, such as lockmgr locks, 358 * simply by holding a critical section. The sequence is: 359 * 360 * (acquire upper level lock) 361 * tsleep_interlock(blah) 362 * (release upper level lock) 363 * tsleep(blah, ...) 364 * 365 * Basically this functions queues us on the tsleep queue without actually 366 * descheduling us. When tsleep() is later called with PINTERLOCK it 367 * assumes the thread was already queued, otherwise it queues it there. 368 * 369 * Thus it is possible to receive the wakeup prior to going to sleep and 370 * the race conditions are covered. 371 */ 372 static __inline void 373 _tsleep_interlock(globaldata_t gd, const volatile void *ident, int flags) 374 { 375 thread_t td = gd->gd_curthread; 376 int id; 377 378 crit_enter_quick(td); 379 if (td->td_flags & TDF_TSLEEPQ) { 380 id = LOOKUP(td->td_wchan); 381 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 382 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) { 383 atomic_clear_cpumask(&slpque_cpumasks[id], 384 gd->gd_cpumask); 385 } 386 } else { 387 td->td_flags |= TDF_TSLEEPQ; 388 } 389 id = LOOKUP(ident); 390 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_sleepq); 391 atomic_set_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 392 td->td_wchan = ident; 393 td->td_wdomain = flags & PDOMAIN_MASK; 394 crit_exit_quick(td); 395 } 396 397 void 398 tsleep_interlock(const volatile void *ident, int flags) 399 { 400 _tsleep_interlock(mycpu, ident, flags); 401 } 402 403 /* 404 * Remove thread from sleepq. Must be called with a critical section held. 405 */ 406 static __inline void 407 _tsleep_remove(thread_t td) 408 { 409 globaldata_t gd = mycpu; 410 int id; 411 412 KKASSERT(td->td_gd == gd); 413 if (td->td_flags & TDF_TSLEEPQ) { 414 td->td_flags &= ~TDF_TSLEEPQ; 415 id = LOOKUP(td->td_wchan); 416 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_sleepq); 417 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 418 atomic_clear_cpumask(&slpque_cpumasks[id], gd->gd_cpumask); 419 td->td_wchan = NULL; 420 td->td_wdomain = 0; 421 } 422 } 423 424 void 425 tsleep_remove(thread_t td) 426 { 427 _tsleep_remove(td); 428 } 429 430 /* 431 * This function removes a thread from the tsleep queue and schedules 432 * it. This function may act asynchronously. The target thread may be 433 * sleeping on a different cpu. 434 * 435 * This function mus be called while in a critical section but if the 436 * target thread is sleeping on a different cpu we cannot safely probe 437 * td_flags. 438 * 439 * This function is only called from a different cpu via setrunnable() 440 * when the thread is in a known sleep. However, multiple wakeups are 441 * possible and we must hold the td to prevent a race against the thread 442 * exiting. 443 */ 444 static __inline 445 void 446 _tsleep_wakeup(struct thread *td) 447 { 448 #ifdef SMP 449 globaldata_t gd = mycpu; 450 451 if (td->td_gd != gd) { 452 lwkt_hold(td); 453 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)tsleep_wakeup_remote, td); 454 return; 455 } 456 #endif 457 _tsleep_remove(td); 458 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 459 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 460 lwkt_schedule(td); 461 } 462 } 463 464 #ifdef SMP 465 static 466 void 467 tsleep_wakeup_remote(struct thread *td) 468 { 469 _tsleep_wakeup(td); 470 lwkt_rele(td); 471 } 472 #endif 473 474 475 /* 476 * General sleep call. Suspends the current process until a wakeup is 477 * performed on the specified identifier. The process will then be made 478 * runnable with the specified priority. Sleeps at most timo/hz seconds 479 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 480 * before and after sleeping, else signals are not checked. Returns 0 if 481 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 482 * signal needs to be delivered, ERESTART is returned if the current system 483 * call should be restarted if possible, and EINTR is returned if the system 484 * call should be interrupted by the signal (return EINTR). 485 * 486 * Note that if we are a process, we release_curproc() before messing with 487 * the LWKT scheduler. 488 * 489 * During autoconfiguration or after a panic, a sleep will simply 490 * lower the priority briefly to allow interrupts, then return. 491 */ 492 int 493 tsleep(const volatile void *ident, int flags, const char *wmesg, int timo) 494 { 495 struct thread *td = curthread; 496 struct lwp *lp = td->td_lwp; 497 struct proc *p = td->td_proc; /* may be NULL */ 498 globaldata_t gd; 499 int sig; 500 int catch; 501 int id; 502 int error; 503 int oldpri; 504 struct callout thandle; 505 506 /* 507 * NOTE: removed KTRPOINT, it could cause races due to blocking 508 * even in stable. Just scrap it for now. 509 */ 510 if (!tsleep_crypto_dump && (tsleep_now_works == 0 || panicstr)) { 511 /* 512 * After a panic, or before we actually have an operational 513 * softclock, just give interrupts a chance, then just return; 514 * 515 * don't run any other procs or panic below, 516 * in case this is the idle process and already asleep. 517 */ 518 splz(); 519 oldpri = td->td_pri; 520 lwkt_setpri_self(safepri); 521 lwkt_switch(); 522 lwkt_setpri_self(oldpri); 523 return (0); 524 } 525 logtsleep2(tsleep_beg, ident); 526 gd = td->td_gd; 527 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 528 529 /* 530 * NOTE: all of this occurs on the current cpu, including any 531 * callout-based wakeups, so a critical section is a sufficient 532 * interlock. 533 * 534 * The entire sequence through to where we actually sleep must 535 * run without breaking the critical section. 536 */ 537 catch = flags & PCATCH; 538 error = 0; 539 sig = 0; 540 541 crit_enter_quick(td); 542 543 KASSERT(ident != NULL, ("tsleep: no ident")); 544 KASSERT(lp == NULL || 545 lp->lwp_stat == LSRUN || /* Obvious */ 546 lp->lwp_stat == LSSTOP, /* Set in tstop */ 547 ("tsleep %p %s %d", 548 ident, wmesg, lp->lwp_stat)); 549 550 /* 551 * We interlock the sleep queue if the caller has not already done 552 * it for us. This must be done before we potentially acquire any 553 * tokens or we can loose the wakeup. 554 */ 555 if ((flags & PINTERLOCKED) == 0) { 556 id = LOOKUP(ident); 557 _tsleep_interlock(gd, ident, flags); 558 } 559 560 /* 561 * Setup for the current process (if this is a process). 562 * 563 * We hold the process token if lp && catch. The resume 564 * code will release it. 565 */ 566 if (lp) { 567 if (catch) { 568 /* 569 * Early termination if PCATCH was set and a 570 * signal is pending, interlocked with the 571 * critical section. 572 * 573 * Early termination only occurs when tsleep() is 574 * entered while in a normal LSRUN state. 575 */ 576 lwkt_gettoken(&p->p_token); 577 if ((sig = CURSIG(lp)) != 0) 578 goto resume; 579 580 /* 581 * Early termination if PCATCH was set and a 582 * mailbox signal was possibly delivered prior to 583 * the system call even being made, in order to 584 * allow the user to interlock without having to 585 * make additional system calls. 586 */ 587 if (p->p_flag & P_MAILBOX) 588 goto resume; 589 590 /* 591 * Causes ksignal to wake us up if a signal is 592 * received (interlocked with p->p_token). 593 */ 594 lp->lwp_flag |= LWP_SINTR; 595 } 596 } else { 597 KKASSERT(p == NULL); 598 } 599 600 /* 601 * Make sure the current process has been untangled from 602 * the userland scheduler and initialize slptime to start 603 * counting. 604 */ 605 if (lp) { 606 p->p_usched->release_curproc(lp); 607 lp->lwp_slptime = 0; 608 } 609 610 /* 611 * If the interlocked flag is set but our cpu bit in the slpqueue 612 * is no longer set, then a wakeup was processed inbetween the 613 * tsleep_interlock() (ours or the callers), and here. This can 614 * occur under numerous circumstances including when we release the 615 * current process. 616 * 617 * Extreme loads can cause the sending of an IPI (e.g. wakeup()'s) 618 * to process incoming IPIs, thus draining incoming wakeups. 619 */ 620 if ((td->td_flags & TDF_TSLEEPQ) == 0) { 621 logtsleep2(ilockfail, ident); 622 goto resume; 623 } 624 625 /* 626 * scheduling is blocked while in a critical section. Coincide 627 * the descheduled-by-tsleep flag with the descheduling of the 628 * lwkt. 629 * 630 * The timer callout is localized on our cpu and interlocked by 631 * our critical section. 632 */ 633 lwkt_deschedule_self(td); 634 td->td_flags |= TDF_TSLEEP_DESCHEDULED; 635 td->td_wmesg = wmesg; 636 637 /* 638 * Setup the timeout, if any. The timeout is only operable while 639 * the thread is flagged descheduled. 640 */ 641 KKASSERT((td->td_flags & TDF_TIMEOUT) == 0); 642 if (timo) { 643 callout_init_mp(&thandle); 644 callout_reset(&thandle, timo, endtsleep, td); 645 } 646 647 /* 648 * Beddy bye bye. 649 */ 650 if (lp) { 651 /* 652 * Ok, we are sleeping. Place us in the SSLEEP state. 653 */ 654 KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0); 655 /* 656 * tstop() sets LSSTOP, so don't fiddle with that. 657 */ 658 if (lp->lwp_stat != LSSTOP) 659 lp->lwp_stat = LSSLEEP; 660 lp->lwp_ru.ru_nvcsw++; 661 lwkt_switch(); 662 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 663 664 /* 665 * And when we are woken up, put us back in LSRUN. If we 666 * slept for over a second, recalculate our estcpu. 667 */ 668 lp->lwp_stat = LSRUN; 669 if (lp->lwp_slptime) 670 p->p_usched->recalculate(lp); 671 lp->lwp_slptime = 0; 672 } else { 673 lwkt_switch(); 674 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 675 } 676 677 /* 678 * Make sure we haven't switched cpus while we were asleep. It's 679 * not supposed to happen. Cleanup our temporary flags. 680 */ 681 KKASSERT(gd == td->td_gd); 682 683 /* 684 * Cleanup the timeout. If the timeout has already occured thandle 685 * has already been stopped, otherwise stop thandle. 686 */ 687 if (timo) { 688 if (td->td_flags & TDF_TIMEOUT) { 689 td->td_flags &= ~TDF_TIMEOUT; 690 error = EWOULDBLOCK; 691 } else { 692 /* does not block when on same cpu */ 693 callout_stop(&thandle); 694 } 695 } 696 697 /* 698 * Make sure we have been removed from the sleepq. In most 699 * cases this will have been done for us already but it is 700 * possible for a scheduling IPI to be in-flight from a 701 * previous tsleep/tsleep_interlock() or due to a straight-out 702 * call to lwkt_schedule() (in the case of an interrupt thread), 703 * causing a spurious wakeup. 704 */ 705 _tsleep_remove(td); 706 td->td_wmesg = NULL; 707 708 /* 709 * Figure out the correct error return. If interrupted by a 710 * signal we want to return EINTR or ERESTART. 711 * 712 * If P_MAILBOX is set no automatic system call restart occurs 713 * and we return EINTR. P_MAILBOX is meant to be used as an 714 * interlock, the user must poll it prior to any system call 715 * that it wishes to interlock a mailbox signal against since 716 * the flag is cleared on *any* system call that sleeps. 717 * 718 * p->p_token is held in the p && catch case. 719 */ 720 resume: 721 if (p) { 722 if (catch && error == 0) { 723 if ((p->p_flag & P_MAILBOX) && sig == 0) { 724 error = EINTR; 725 } else if (sig != 0 || (sig = CURSIG(lp))) { 726 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 727 error = EINTR; 728 else 729 error = ERESTART; 730 } 731 } 732 if (catch) 733 lwkt_reltoken(&p->p_token); 734 lp->lwp_flag &= ~(LWP_BREAKTSLEEP | LWP_SINTR); 735 p->p_flag &= ~P_MAILBOX; 736 } 737 logtsleep1(tsleep_end); 738 crit_exit_quick(td); 739 return (error); 740 } 741 742 /* 743 * Interlocked spinlock sleep. An exclusively held spinlock must 744 * be passed to ssleep(). The function will atomically release the 745 * spinlock and tsleep on the ident, then reacquire the spinlock and 746 * return. 747 * 748 * This routine is fairly important along the critical path, so optimize it 749 * heavily. 750 */ 751 int 752 ssleep(const volatile void *ident, struct spinlock *spin, int flags, 753 const char *wmesg, int timo) 754 { 755 globaldata_t gd = mycpu; 756 int error; 757 758 _tsleep_interlock(gd, ident, flags); 759 spin_unlock_quick(gd, spin); 760 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 761 spin_lock_quick(gd, spin); 762 763 return (error); 764 } 765 766 int 767 lksleep(const volatile void *ident, struct lock *lock, int flags, 768 const char *wmesg, int timo) 769 { 770 globaldata_t gd = mycpu; 771 int error; 772 773 _tsleep_interlock(gd, ident, flags); 774 lockmgr(lock, LK_RELEASE); 775 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 776 lockmgr(lock, LK_EXCLUSIVE); 777 778 return (error); 779 } 780 781 /* 782 * Interlocked mutex sleep. An exclusively held mutex must be passed 783 * to mtxsleep(). The function will atomically release the mutex 784 * and tsleep on the ident, then reacquire the mutex and return. 785 */ 786 int 787 mtxsleep(const volatile void *ident, struct mtx *mtx, int flags, 788 const char *wmesg, int timo) 789 { 790 globaldata_t gd = mycpu; 791 int error; 792 793 _tsleep_interlock(gd, ident, flags); 794 mtx_unlock(mtx); 795 error = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 796 mtx_lock_ex_quick(mtx, wmesg); 797 798 return (error); 799 } 800 801 /* 802 * Interlocked serializer sleep. An exclusively held serializer must 803 * be passed to zsleep(). The function will atomically release 804 * the serializer and tsleep on the ident, then reacquire the serializer 805 * and return. 806 */ 807 int 808 zsleep(const volatile void *ident, struct lwkt_serialize *slz, int flags, 809 const char *wmesg, int timo) 810 { 811 globaldata_t gd = mycpu; 812 int ret; 813 814 ASSERT_SERIALIZED(slz); 815 816 _tsleep_interlock(gd, ident, flags); 817 lwkt_serialize_exit(slz); 818 ret = tsleep(ident, flags | PINTERLOCKED, wmesg, timo); 819 lwkt_serialize_enter(slz); 820 821 return ret; 822 } 823 824 /* 825 * Directly block on the LWKT thread by descheduling it. This 826 * is much faster then tsleep(), but the only legal way to wake 827 * us up is to directly schedule the thread. 828 * 829 * Setting TDF_SINTR will cause new signals to directly schedule us. 830 * 831 * This routine must be called while in a critical section. 832 */ 833 int 834 lwkt_sleep(const char *wmesg, int flags) 835 { 836 thread_t td = curthread; 837 int sig; 838 839 if ((flags & PCATCH) == 0 || td->td_lwp == NULL) { 840 td->td_flags |= TDF_BLOCKED; 841 td->td_wmesg = wmesg; 842 lwkt_deschedule_self(td); 843 lwkt_switch(); 844 td->td_wmesg = NULL; 845 td->td_flags &= ~TDF_BLOCKED; 846 return(0); 847 } 848 if ((sig = CURSIG(td->td_lwp)) != 0) { 849 if (SIGISMEMBER(td->td_proc->p_sigacts->ps_sigintr, sig)) 850 return(EINTR); 851 else 852 return(ERESTART); 853 854 } 855 td->td_flags |= TDF_BLOCKED | TDF_SINTR; 856 td->td_wmesg = wmesg; 857 lwkt_deschedule_self(td); 858 lwkt_switch(); 859 td->td_flags &= ~(TDF_BLOCKED | TDF_SINTR); 860 td->td_wmesg = NULL; 861 return(0); 862 } 863 864 /* 865 * Implement the timeout for tsleep. 866 * 867 * We set LWP_BREAKTSLEEP to indicate that an event has occured, but 868 * we only call setrunnable if the process is not stopped. 869 * 870 * This type of callout timeout is scheduled on the same cpu the process 871 * is sleeping on. Also, at the moment, the MP lock is held. 872 */ 873 static void 874 endtsleep(void *arg) 875 { 876 thread_t td = arg; 877 struct lwp *lp; 878 879 crit_enter(); 880 881 /* 882 * Do this before we potentially block acquiring the token. Setting 883 * TDF_TIMEOUT tells tsleep that we have already stopped the callout. 884 */ 885 lwkt_hold(td); 886 td->td_flags |= TDF_TIMEOUT; 887 888 /* 889 * This can block 890 */ 891 if ((lp = td->td_lwp) != NULL) 892 lwkt_gettoken(&lp->lwp_proc->p_token); 893 894 /* 895 * Only do nominal wakeup processing if TDF_TIMEOUT and 896 * TDF_TSLEEP_DESCHEDULED are both still set. Otherwise 897 * we raced a wakeup or we began executed and raced due to 898 * blocking in the token above, and should do nothing. 899 */ 900 if ((td->td_flags & (TDF_TIMEOUT | TDF_TSLEEP_DESCHEDULED)) == 901 (TDF_TIMEOUT | TDF_TSLEEP_DESCHEDULED)) { 902 if (lp) { 903 lp->lwp_flag |= LWP_BREAKTSLEEP; 904 if (lp->lwp_proc->p_stat != SSTOP) 905 setrunnable(lp); 906 } else { 907 _tsleep_wakeup(td); 908 } 909 } 910 if (lp) 911 lwkt_reltoken(&lp->lwp_proc->p_token); 912 lwkt_rele(td); 913 crit_exit(); 914 } 915 916 /* 917 * Make all processes sleeping on the specified identifier runnable. 918 * count may be zero or one only. 919 * 920 * The domain encodes the sleep/wakeup domain AND the first cpu to check 921 * (which is always the current cpu). As we iterate across cpus 922 * 923 * This call may run without the MP lock held. We can only manipulate thread 924 * state on the cpu owning the thread. We CANNOT manipulate process state 925 * at all. 926 * 927 * _wakeup() can be passed to an IPI so we can't use (const volatile 928 * void *ident). 929 */ 930 static void 931 _wakeup(void *ident, int domain) 932 { 933 struct tslpque *qp; 934 struct thread *td; 935 struct thread *ntd; 936 globaldata_t gd; 937 #ifdef SMP 938 cpumask_t mask; 939 #endif 940 int id; 941 942 crit_enter(); 943 logtsleep2(wakeup_beg, ident); 944 gd = mycpu; 945 id = LOOKUP(ident); 946 qp = &gd->gd_tsleep_hash[id]; 947 restart: 948 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 949 ntd = TAILQ_NEXT(td, td_sleepq); 950 if (td->td_wchan == ident && 951 td->td_wdomain == (domain & PDOMAIN_MASK) 952 ) { 953 KKASSERT(td->td_gd == gd); 954 _tsleep_remove(td); 955 if (td->td_flags & TDF_TSLEEP_DESCHEDULED) { 956 td->td_flags &= ~TDF_TSLEEP_DESCHEDULED; 957 lwkt_schedule(td); 958 if (domain & PWAKEUP_ONE) 959 goto done; 960 } 961 goto restart; 962 } 963 } 964 965 #ifdef SMP 966 /* 967 * We finished checking the current cpu but there still may be 968 * more work to do. Either wakeup_one was requested and no matching 969 * thread was found, or a normal wakeup was requested and we have 970 * to continue checking cpus. 971 * 972 * It should be noted that this scheme is actually less expensive then 973 * the old scheme when waking up multiple threads, since we send 974 * only one IPI message per target candidate which may then schedule 975 * multiple threads. Before we could have wound up sending an IPI 976 * message for each thread on the target cpu (!= current cpu) that 977 * needed to be woken up. 978 * 979 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 980 * should be ok since we are passing idents in the IPI rather then 981 * thread pointers. 982 */ 983 if ((domain & PWAKEUP_MYCPU) == 0 && 984 (mask = slpque_cpumasks[id] & gd->gd_other_cpus) != 0) { 985 lwkt_send_ipiq2_mask(mask, _wakeup, ident, 986 domain | PWAKEUP_MYCPU); 987 } 988 #endif 989 done: 990 logtsleep1(wakeup_end); 991 crit_exit(); 992 } 993 994 /* 995 * Wakeup all threads tsleep()ing on the specified ident, on all cpus 996 */ 997 void 998 wakeup(const volatile void *ident) 999 { 1000 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 1001 } 1002 1003 /* 1004 * Wakeup one thread tsleep()ing on the specified ident, on any cpu. 1005 */ 1006 void 1007 wakeup_one(const volatile void *ident) 1008 { 1009 /* XXX potentially round-robin the first responding cpu */ 1010 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 1011 } 1012 1013 /* 1014 * Wakeup threads tsleep()ing on the specified ident on the current cpu 1015 * only. 1016 */ 1017 void 1018 wakeup_mycpu(const volatile void *ident) 1019 { 1020 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 1021 } 1022 1023 /* 1024 * Wakeup one thread tsleep()ing on the specified ident on the current cpu 1025 * only. 1026 */ 1027 void 1028 wakeup_mycpu_one(const volatile void *ident) 1029 { 1030 /* XXX potentially round-robin the first responding cpu */ 1031 _wakeup(__DEALL(ident), PWAKEUP_MYCPU|PWAKEUP_ONE); 1032 } 1033 1034 /* 1035 * Wakeup all thread tsleep()ing on the specified ident on the specified cpu 1036 * only. 1037 */ 1038 void 1039 wakeup_oncpu(globaldata_t gd, const volatile void *ident) 1040 { 1041 #ifdef SMP 1042 if (gd == mycpu) { 1043 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 1044 } else { 1045 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), PWAKEUP_MYCPU); 1046 } 1047 #else 1048 _wakeup(__DEALL(ident), PWAKEUP_MYCPU); 1049 #endif 1050 } 1051 1052 /* 1053 * Wakeup one thread tsleep()ing on the specified ident on the specified cpu 1054 * only. 1055 */ 1056 void 1057 wakeup_oncpu_one(globaldata_t gd, const volatile void *ident) 1058 { 1059 #ifdef SMP 1060 if (gd == mycpu) { 1061 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1062 } else { 1063 lwkt_send_ipiq2(gd, _wakeup, __DEALL(ident), 1064 PWAKEUP_MYCPU | PWAKEUP_ONE); 1065 } 1066 #else 1067 _wakeup(__DEALL(ident), PWAKEUP_MYCPU | PWAKEUP_ONE); 1068 #endif 1069 } 1070 1071 /* 1072 * Wakeup all threads waiting on the specified ident that slept using 1073 * the specified domain, on all cpus. 1074 */ 1075 void 1076 wakeup_domain(const volatile void *ident, int domain) 1077 { 1078 _wakeup(__DEALL(ident), PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 1079 } 1080 1081 /* 1082 * Wakeup one thread waiting on the specified ident that slept using 1083 * the specified domain, on any cpu. 1084 */ 1085 void 1086 wakeup_domain_one(const volatile void *ident, int domain) 1087 { 1088 /* XXX potentially round-robin the first responding cpu */ 1089 _wakeup(__DEALL(ident), 1090 PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 1091 } 1092 1093 /* 1094 * setrunnable() 1095 * 1096 * Make a process runnable. lp->lwp_proc->p_token must be held on call. 1097 * This only has an effect if we are in SSLEEP. We only break out of the 1098 * tsleep if LWP_BREAKTSLEEP is set, otherwise we just fix-up the state. 1099 * 1100 * NOTE: With p_token held we can only safely manipulate the process 1101 * structure and the lp's lwp_stat. 1102 */ 1103 void 1104 setrunnable(struct lwp *lp) 1105 { 1106 ASSERT_LWKT_TOKEN_HELD(&lp->lwp_proc->p_token); 1107 crit_enter(); 1108 if (lp->lwp_stat == LSSTOP) 1109 lp->lwp_stat = LSSLEEP; 1110 if (lp->lwp_stat == LSSLEEP && (lp->lwp_flag & LWP_BREAKTSLEEP)) 1111 _tsleep_wakeup(lp->lwp_thread); 1112 crit_exit(); 1113 } 1114 1115 /* 1116 * The process is stopped due to some condition, usually because p_stat is 1117 * set to SSTOP, but also possibly due to being traced. 1118 * 1119 * NOTE! If the caller sets SSTOP, the caller must also clear P_WAITED 1120 * because the parent may check the child's status before the child actually 1121 * gets to this routine. 1122 * 1123 * This routine is called with the current lwp only, typically just 1124 * before returning to userland. 1125 * 1126 * Setting LWP_BREAKTSLEEP before entering the tsleep will cause a passive 1127 * SIGCONT to break out of the tsleep. 1128 */ 1129 void 1130 tstop(void) 1131 { 1132 struct lwp *lp = curthread->td_lwp; 1133 struct proc *p = lp->lwp_proc; 1134 struct proc *q; 1135 1136 crit_enter(); 1137 /* 1138 * If LWP_WSTOP is set, we were sleeping 1139 * while our process was stopped. At this point 1140 * we were already counted as stopped. 1141 */ 1142 if ((lp->lwp_flag & LWP_WSTOP) == 0) { 1143 /* 1144 * If we're the last thread to stop, signal 1145 * our parent. 1146 */ 1147 p->p_nstopped++; 1148 lp->lwp_flag |= LWP_WSTOP; 1149 wakeup(&p->p_nstopped); 1150 if (p->p_nstopped == p->p_nthreads) { 1151 /* 1152 * Token required to interlock kern_wait() 1153 */ 1154 q = p->p_pptr; 1155 PHOLD(q); 1156 lwkt_gettoken(&q->p_token); 1157 p->p_flag &= ~P_WAITED; 1158 wakeup(p->p_pptr); 1159 if ((q->p_sigacts->ps_flag & PS_NOCLDSTOP) == 0) 1160 ksignal(q, SIGCHLD); 1161 lwkt_reltoken(&q->p_token); 1162 PRELE(q); 1163 } 1164 } 1165 while (p->p_stat == SSTOP) { 1166 lp->lwp_flag |= LWP_BREAKTSLEEP; 1167 lp->lwp_stat = LSSTOP; 1168 tsleep(p, 0, "stop", 0); 1169 } 1170 p->p_nstopped--; 1171 lp->lwp_flag &= ~LWP_WSTOP; 1172 crit_exit(); 1173 } 1174 1175 /* 1176 * Compute a tenex style load average of a quantity on 1177 * 1, 5 and 15 minute intervals. 1178 */ 1179 static int loadav_count_runnable(struct lwp *p, void *data); 1180 1181 static void 1182 loadav(void *arg) 1183 { 1184 struct loadavg *avg; 1185 int i, nrun; 1186 1187 nrun = 0; 1188 alllwp_scan(loadav_count_runnable, &nrun); 1189 avg = &averunnable; 1190 for (i = 0; i < 3; i++) { 1191 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1192 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1193 } 1194 1195 /* 1196 * Schedule the next update to occur after 5 seconds, but add a 1197 * random variation to avoid synchronisation with processes that 1198 * run at regular intervals. 1199 */ 1200 callout_reset(&loadav_callout, hz * 4 + (int)(krandom() % (hz * 2 + 1)), 1201 loadav, NULL); 1202 } 1203 1204 static int 1205 loadav_count_runnable(struct lwp *lp, void *data) 1206 { 1207 int *nrunp = data; 1208 thread_t td; 1209 1210 switch (lp->lwp_stat) { 1211 case LSRUN: 1212 if ((td = lp->lwp_thread) == NULL) 1213 break; 1214 if (td->td_flags & TDF_BLOCKED) 1215 break; 1216 ++*nrunp; 1217 break; 1218 default: 1219 break; 1220 } 1221 lwkt_yield(); 1222 return(0); 1223 } 1224 1225 /* ARGSUSED */ 1226 static void 1227 sched_setup(void *dummy) 1228 { 1229 callout_init_mp(&loadav_callout); 1230 callout_init_mp(&schedcpu_callout); 1231 1232 /* Kick off timeout driven events by calling first time. */ 1233 schedcpu(NULL); 1234 loadav(NULL); 1235 } 1236 1237